Need ArXiv endorsement.. CS. Multiagent-Systems [pdf] (www.socialcompute.dev)

A new working paper by Marcos Pazzarelli proposes an innovative architecture for social simulations where large language models (LLMs) serve as stochastic proposal engines under a deterministic control system. This architecture enforces a strict separation between proposal generation and state alteration by having all agent actions go through a deterministic validation process that ensures suggested behaviors comply with predefined rules and constraints. The simulation operates in discrete time ticks, guiding agents through phases of perception, proposal, validation, and resolution, significantly enhancing the consistency and reliability of simulation outcomes. This approach is noteworthy for the AI/ML community as it addresses common failures observed in traditional LLM-driven simulations, such as state drift and unsupported assertiveness, which can undermine the integrity of the simulation. By defining determinism at multiple operational levels, the paper aims to establish a framework that enhances epistemic integrity, observability, and long-horizon coherence, while retaining the beneficial variability of agent behavior. The detailed design includes structured state representations, telemetry for debugging, and mechanisms for handling unsupported claims, positioning this work as a potential advancement in the development of robust and interpretable multi-agent systems.